(Deno) LTX Model Loader

LTX 2.3 is 22B of synchronized video-plus-audio, and there are three different ways people load it into ComfyUI: the full checkpoint, the split KJ-style diffusion/VAE files, or GGUF quantizations for smaller VRAM. Each needs a different set of loader nodes, different files, and a different mental model. (Deno) LTX Model Loader collapses all three into one node with a pipeline_mode switch, and outputs the same four things every time: model, clip, video_vae, and audio_vae.

The modes

• Checkpoint Style - the full LTX 2.3 checkpoint (ltx-2.3-22b-dev / distilled) through ComfyUI's built-in checkpoint loading. Simplest, but it's a big file and the most VRAM-hungry path.
• KJ Style - split loading the way Kijai's workflow does it: separate diffusion model, video VAE, audio VAE, text encoder, and text projection. More moving parts, but it's the path that lets you mix quantized weights and keep VRAM sane.
• GGUF Style - LTX 2.3 GGUFs through ComfyUI-GGUF's UNet loading. This is how people actually run 2.3 on 8–16GB cards; the official model card's 32GB "minimum" is fiction for quantized users.

The clip output is built from the Gemma 3 12B text encoder via LTXAVTextEncoderLoader, with clip_device and weight_dtype set to default - leave them there unless you know why you're changing them.

Why LTX specifically needs this

LTX 2.3 splits its VAE. Video and audio each have their own VAE, and you need both to generate sound-synchronized video - one video_vae and one audio_vae output. On the KJ/GGUF styles the node leans on KJNodes' VAELoaderKJ for the split VAEs, and it includes an audio-VAE compatibility fallback for mixed ComfyUI/KJNodes environments. So the node isn't just a menu; it's making sure the video VAE and audio VAE actually come from the same generation of the model, which is where half of "my LTX 2.3 is broken" reports start.

Dependencies and gotchas

The pack itself installs normally:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI

But this node is a workflow helper on top of other people's engines, and it's upfront about it. GGUF Style requires ComfyUI-GGUF installed from Manager, plus a ComfyUI core new enough to carry diffusion-model metadata. KJ/GGUF styles require comfyui-kjnodes for the VAE loaders. If a mode is missing its backend, you get a clear one-line error naming the missing pack rather than a stack trace - the error strings in the source literally tell you to install ComfyUI-GGUF or kjnodes and restart.

The trap to avoid: loading the checkpoint in Checkpoint Style and assuming the audio VAE came along. Checkpoint files don't always carry the split audio VAE the way the standalone loaders do. If your workflow stalls on audio, switch to KJ Style with the explicit audio_vae_name populated. And if you're on a 12–16GB card, go GGUF Style first - that's the path the pack's own Easy Model Download Helper presets are built around.